Can India Capitalise on Beijing’s BRICS AI Pitch?

The most consequential technology imports of the coming decade may never pass through a port. They will arrive as downloads: model weights, the long lists of numbers an AI system learns during training, and all that is needed to run it. Anyone with the file can modify it, fine-tune it (retrain it on their own data), and run it locally. That changes both the economics and the politics of dependence.

At the second session of the 18^th BRICS (Brazil, Russia, India, China, South Africa) Summit in New Delhi on 13 September 2026, Xi Jinping proposed an “open source and inclusive AI initiative”. China, he said, would be a pioneer in establishing a BRICS AI open-source community, support cooperation on large language models (LLMs, the technology behind chatbots), run specialised training, and build an open AI ecosystem.

His other proposals went further. They included a BRICS digital ecosystem cloud platform, help with smart factories and their “standards and norms”, and an alliance to train engineers and mutually recognise their competencies. Read together, this is not an offer of models alone: model, cloud, factory standard, and engineer’s certificate are layers of a technology stack, that is, the chain on which AI runs, with chips at the bottom, then the data centres that house them, then models, then the applications people actually use.

Open-weight models — those whose weights are published for download, even when the training data and code are not — lower the cost of sophisticated AI and let firms and governments adapt it to local languages, data, and uses. But openness at the model layer does not decide who controls the hardware, cloud, standards, and institutions through which models are deployed.

India’s position is unusual. In July, 29 countries signed an agreement in Shanghai creating the World Artificial Intelligence Cooperation Organization (WAICO), a China-led body for coordinating AI cooperation among its members; India was not among them. In February, India joined Pax Silica, the US-led initiative spanning critical minerals, chip fabrication, and AI infrastructure.

New Delhi then chaired the summit where Beijing made its pitch, which the New Delhi Declaration did not take up. For now, it remains a Chinese proposal, and China hosts BRICS in 2027. The real question is therefore less “Chinese or American technology?” than how dependence is distributed across the stack.

The economics of open models

The pull of Chinese open models is real and should not be reduced to geopolitics. Hugging Face, the main public repository where AI models are shared, reported that Chinese models took about 41% of downloads on its platform in the year to spring 2026. Of 178 Chinese releases above 20 billion parameters (the adjustable numbers inside a model, a rough proxy for its size) this year, 59% carried the Apache 2.0 license and 22% MIT (Massachusetts Institute of Technology) licences, both of which permit commercial reuse.

Hugging Face cautions that downloads measure activity on its platform, not market share or quality. On price, the Chinese developer DeepSeek’s V4 Pro lists output at 87 cents per million tokens (chunks of text, roughly three-quarters of a word each), though per-task comparisons narrow the gap with American models.

For Indian firms, price is not the whole story. Downloadable weights let developers fine-tune for Indian languages, run systems on infrastructure they control, and experiment without permanent reliance on someone else’s API (application programming interface). More than 60% of Indian start-ups seeking subsidised GPUs (graphics processing unit) want them for inference (running finished models to answer queries, rather than the far costlier work of training new ones), and many adapt existing open models rather than train their own.

Openness can even aid defence. In July, when closed frontier models (the most advanced systems, usable only through their makers’ own services) declined to help Hugging Face analyse code from an intrusion, its team finished the job on an open model running on its own servers.

Open models can support autonomy, but they do not suspend commercial strategy. When weights are given away, value migrates to cloud services, interfaces, hardware, developer ecosystems, and standards. Hugging Face notes this shift, along with tighter terms on some very large recent models and the growing optimisation of Chinese models for China’s own chips.

Hence the key distinction: the openness of an artefact versus the openness of an ecosystem. A permissively licensed model can still sit inside an environment whose hardware, tools, cloud, or standards are controlled elsewhere.

The manufacturing analogy

India’s manufacturing experience is a useful, if imperfect, guide. Cheap Chinese components cut costs from electronics to solar and helped Indian factories export. They also created a particular shape of dependence.

India’s trade deficit with China reached US$112.1 billion in 2025-26, US$43.1 billion of it in electrical machinery. China supplied at least 80% of imports in 71 electronics product lines, up from 44 in 2018-19. In solar, India’s 233 GW (gigawatt) of module capacity is about seven times its cell capacity and 116 times its ingot and wafer capacity, the stage where raw polysilicon is cast and sliced into the discs that become solar cells.

Call it the “inverted-pyramid problem”: a broad base of downstream assembly resting on a narrow tip of upstream capability. AI could reproduce that shape, with a vast community of users, app builders, and fine-tuning firms resting on processors, foundation models (the general-purpose systems others build on), the software toolkits that run them, cloud capacity, and standards supplied from elsewhere. That would still create value. It would not be technological independence.

China’s own rise teaches the other half of the lesson. It came less from replacing imports than from dense supplier networks, engineering talent, patient finance, and ferocious domestic competition. Its AI sector shows the same reflexes: Baidu went from no model releases on Hugging Face in 2024 to more than 100 in 2025.

China also climbed a ladder others left open; the shift away from Meta’s Llama, until then the most widely used open model family, began with Alibaba’s Qwen 2.5 in 2024. Now the ladder runs both ways, with some large American open releases this year built on Chinese models.

Adopting foreign technology and building capability are not opposites. Dependence becomes consequential when imports are consumed without building the skills, institutions and upstream capacity to modify, reproduce or eventually replace them.

Dependence across the AI stack

Treat AI as a layered system and the problem sharpens. At the base, the IndiaAI Mission has onboarded more than 38,000 GPUs at a subsidised Rs. 65 an hour, but advanced processors remain overwhelmingly foreign designed. India’s first commercial fab (chip fabrication plant) at Dholera targets the 28-nanometre node – a mature generation of chipmaking used in cars and industrial electronics – not AI accelerators, the specialised chips that train and run large models.

Above hardware sit cloud platforms, the software that trains and runs models, the file formats they come in, the weights themselves, the datasets they learn from, the tests used to score them, and applications. The risks differ by layer: imported processors create supply-chain exposure; foreign cloud raises questions of jurisdiction and operational control; external model families shape the tools and conventions developers adopt; imported benchmarks (the standard tests by which models are ranked) shape what counts as quality or safety.

Model behaviour adds a contested dimension. A September 2025 evaluation by the US Center for AI Standards and Innovation (CAISI) reported that DeepSeek models echoed inaccurate Chinese Communist Party narratives four times as often as American reference models. It also found agents — AI systems left to act on their own, such as browsing the web or running code — built on DeepSeek’s most secure model were 12 times likelier to follow hijacking instructions (malicious commands hidden in the material they read).

CAISI is a US government body tasked with assessing Chinese models, so its findings merit independent replication rather than reflexive acceptance or dismissal. Open weights make such replication possible in a way closed APIs — services you can only send questions to, not inspect — rarely do. But access is not interpretability, that is, being able to tell why a model does what it does: research on “sleeper agent” models (systems with harmful behaviour deliberately planted in them) shows hidden behaviour can survive safety training, and hundreds of billions of parameters are not transparent the way source code is.

An Indian developer running a Chinese-origin model on American-designed chips with internationally built software shows why AI resists a single national label. The policy question is which interdependencies are manageable, and which capabilities matter enough to keep under domestic control.

Standards as infrastructure

This is why standards deserve more attention than they get. Both superpowers now treat AI as a stack. China’s proposals bundle models with cloud, manufacturing standards and training. The US Commerce Department has received 78 consortium applications to export full-stack American AI packages spanning hardware, data, models, cybersecurity, and applications.

The rulebooks differ. China has enforced binding generative-AI rules since August 2023 and mandatory labelling of AI content since September 2025. The EU deferred its high-risk AI obligations to December 2027, partly because harmonised standards were not ready. India’s November 2025 guidelines rely on existing laws rather than a dedicated AI statute, and its data-protection law applies fully only from May 2027.

Standards outlast products. In telecommunications, specifications, and patent arrangements shaped procurement across generations. In AI, similar influence may run through model formats, benchmarks, cloud interfaces, safety tests, provenance standards, and developer tools.

The question is not only whose models India uses, but whether Indian institutions help write the rules for evaluating, exchanging, and deploying them. The New Delhi Declaration already provides for a BRICS technical session on AI standardisation.

What BRICS could guarantee

BRICS is useful here precisely because it is heterogeneous. Its foreign ministers failed to agree a joint statement in May; Brazil, Russia and Indonesia joined WAICO; the UAE belongs to both BRICS and Pax Silica. That can stall agreement, but it also rewards interoperability (the ability of systems from different suppliers to work together) over exclusivity.

Members also have a text to hold one another to. The New Delhi Declaration commits them to implement the 2025 BRICS Leaders’ Statement on AI governance. That statement encourages open foundational models, mechanisms for researchers to scrutinise and audit AI systems, and inclusive standard-setting for interoperability.

Turned into conditions, any shared BRICS arrangement would need: published model documentation and training-data summaries; independent security testing before adoption; signed, provenance-tracked model files (so users can verify where a file came from and that it has not been altered) and incident reporting; audit rights, with data kept under national law; portability, so any member can host a model at home and walk away.

Data, language, and local capability

The strongest case for cooperation lies where no single foreign provider can solve India’s problems, and language is the clearest example. BHASHINI, the government’s language-data platform, gathers voice, text and translations across 22 Indian languages, and the Indian start-up Sarvam released open models of 30 billion and 105 billion parameters in February.

Yet compute, meaning the raw computing power to train and run models, cannot fix a data gap. Omprakash Subbarao of the research arm of the Indian Institute of Science (IISc) notes that only a small share of the IndiaAI budget goes to datasets and warns that more GPUs cannot offset a thin Indic training corpus, the body of text a model learns from.

That changes the economics of sovereignty. An independent frontier-model ecosystem needs enormous capital. Stronger datasets, evaluation institutions and multilingual research need far less, and they let India make better use of models built anywhere.

A BRICS multilingual data commons — a shared pool of language data — could help, provided consent, ownership, localisation, benchmark control, and intellectual property are settled before the data moves.

Four possible configurations

India’s options fall into four broad configurations, each with its own dependencies:

• Extensive participation in shared BRICS models, training and infrastructure would cut costs and speed access but could deepen reliance on standards and infrastructure set by the largest contributor.
• Narrow participation in research, datasets, benchmarking and evaluation limits infrastructural dependence, but works only if Indian institutions can independently test models and enforce interoperability.
• A predominantly domestic stack maximises control over selected layers, at high fiscal cost, over long timelines, and with the risk that protected systems end up costlier or weaker.
• Deliberate interoperability, running hardware, models, and software from several ecosystems under Indian evaluation, data governance, and security testing, avoids reliance on any one system but demands considerable technical and regulatory capacity.

These are not mutually exclusive. Different layers may justify different mixes of national control, cooperation and commercial dependence.

Access is not capability

The central distinction is not between openness and closure, or between Chinese and American technology. It is between access and capability. Open models broaden access, which matters for countries and firms that cannot finance frontier development, but cheap access does not by itself create the capacity to understand, modify, evaluate, or reproduce the technology.

Nor is strategic autonomy the same as autarky. An Indian model with weak security is no safer than an audited import, and a domestic chip at triple the cost is a tax, not sovereignty. A foreign system need not create harmful dependence if it can be independently tested, substituted, and integrated without surrendering critical functions.

What is significant about China’s BRICS proposal, then, is not its willingness to share models. It is the prospect that models, infrastructure, standards and training could develop as one ecosystem. For India, the consequences would depend less on how many models it downloads than on what capabilities it builds around them.

Industrial history suggests cheap foreign technology is transformative when it accelerates domestic learning and upgrading. It does far less when an economy mainly gets good at using technologies whose upstream layers lie beyond its control. AI compresses that lesson into a much shorter cycle.

Model weights cross borders almost instantly. Scientific capability, engineering depth, trusted institutions, specialised datasets, and standards do not; they must be accumulated.

The question for India is not whether foreign open models should enter its ecosystem; many already have. It is whether they enlarge India’s capacity to evaluate, adapt and eventually originate advanced AI, or simply enlarge the list of technologies on which it depends.

Pradeep Kumar is a data scientist and researcher working at the intersection of machine learning, natural language processing, generative AI, and public policy. Trained as an electrical engineer at NIT Jamshedpur and TU Eindhoven, he began his career in industrial automation and robotics at Tata Motors and Maruti Suzuki. Views expressed are personal.

This piece first appeared on Ideas for India and has been republished under an arrangement.

The Wire is now on WhatsApp. Follow our channel for sharp analysis and opinions on the latest developments.